Chapter 2. In R.S. Michalskl,
26
J.G. Carbonell and T.M. Mitchell
(Eds.), tochine learning, an
artificial approach. Palo Alto,
CA: Tioga Publishing Co., 1983.
Bib. 0552.
WHY SHOULD
MACHINES LEARN?
Herbert A. Simon
Carnegie-Melton University
2.1 INTRODUCTION
When I agreed to write this chapter, I thought I could simply expand a
paper that I wrote for the Carnegie Symposium on Cognition, since the topic of
that symposium was also learning. The difficulty with plagiarizing that paper is
that it was really about psychology, whereas this book is concerned with
machine learning. Now although we all believe machines can simulate human
thought unless we're vitalists, and there aren't any of those around any
more still, I didn't think that was what was intended by the title of the book. I
didn't think it was appropriate to write about psychology.
When my chapter finally was outlined and written, it surprised me a bit;
whether it will surprise you or not, we can leave, to the event. My chapter
turned out to propose a thesis to which perhaps the other chapters in this volume
will serve as antitheses. That will allow us to arrive at the great Hegelian syn-
thesis that we all wish for.
2.2 HUMAN LEARNING AND MACHINE LEARNING
I must begin, after all, by saying something about human teaming, because
I want to compare and contrast what is involved in human learning with what is
involved in machine learning. Out of the synthesis of that contrast in itself a
thesis and antithesis will come my thesis.
CHAPTER 2: WHY SHOULD MACHINES LEARN?
2.2.1 Tedlousness of Human Learning
The first obvious fact about human learning is that it's horribly slow. It
takes decades for human beings to learn anything. It took all of us six years just
to get up to starting speed for school, and then twenty more years to become
cognitive scientists or computer scientists. That is the minimum some of us
took even longer than that. So, we're terribly slow learners. We maintain big
expensive educational systems that arc supposed to make the process effective,
but with all we've been able to do with them to say nothing of computer aided
instruction it remains a terribly slow process.
I can still remember, although it was 45 years ago, trying to learn how to
do multiple regressions by the Gauss-Doolinle method with the aid of a desk
calculator. There's nothing complicated about the method except when you're
learning it. And then it seems terribly mysterious. You wonder why this gets
multiplied by that, and after a long while it gradually dawns on you. As a mat-
ter of fact you can carry out the calculations long before you understand the
rationale for the procedure.
Learning the linear programming simplex method also illustrates this point
and another one as well. Even after you've learned it. even after you've under-
stood it, even after (in principle) you can do it, you still can't really do it be-
cause you can't compute fast enough. I don't know of any humans who cal-
culate solutions of LP problems by the simplex method; as fur as I know it's all
done by computers. The human doesn't even have to know the simplex method;
he just has to know the program library cookbook statistics, or cookbook com-
puting, which we all do most of the time.
So human learning is a long, slow process. It should give us some pause,
when we build machine learning systems, to imagine what can possibly be going
on during all the time a human being is mastering a "simple" skill. We should
ask whether we really want to make the computer go through that tedious
process, or whether machines should be programmed directly to perform tasks,
avoiding humanoid learning entirely.
Of course we might discover a trick: a method of machine learning that
was orders of magnitude faster than human learning. Whether such tricks exist
depend on whether the inefficiencies of human learning derive from peculiar
proper lies of the human information processing system or whether they will be
present in any system that tries to extract patterns or other kinds of information
from complex, noisy situations and to retain those patterns in a manner that
makes them available for later use. The search for such tricks that manage to
escape the tcdiousncss of human learning, however, provides a strong motivation
for research in machine learning.
SIMON .27
2.2.2 Learning and Copying
The second distinctive feature about human learning is that there's no copy
process. In contrast, once you get a debugged program in the computer you can
have as many copies as you want (given equivalence in operating systems and
hardware). You can have these copies free, or almost free. When one computer
has learned it, they've all learned it in principle. An algorithm only has to be
invented once not a billion times. .
I've been involved a little bit in tutoring someone during the last few
weeks in beginning calculus. I think I know the calculus pretty well 1 haven't
used it much for years, but it comes back. Yet I find it terribly frustrating trying
to transfer my knowledge and skill to another human head. I'd like to open the
lid and stuff the program in. But for one thing, I don't know where it is in my
head, I don't even know what language it's encoded in. For another thing, I
have no way of transferring it to the other head. That, of course, is why we
humans go through the slow learning process because we can't copy and trans-
fer programs.
2.2.3 Why Machine Learning?
Contrast this with the machine learning task. In machine learning, the
minute you have the debugged program you read it into the computer and it
runs. The computer docs what the psychologists call "one-trial learning". And,
as I've already indicated, what is learned can be copied ad nauseam. So, if one
thinks about that a little, one says, "What's all this about machine learning?
Why arc we interested in it if by machine learning we mean anything that's at
all like human learning? Who what madman would put a computer through
twenty years of hard labor to make a cognitive scientist or a computer scientist
out of it? Let's forget this nonsense just program it." It would appear that,
now that we have computers, the whole topic of learning has become just one
grand irrelevancy for computer science.
I have already qualified that conclusion in one respect: we do have reason
to search for machine learning programs that will avoid the inefficiencies of
human learning, although we must be alert to the possibility that such programs
cannot, in principle, be constructed. The difficulty may be intrinsic in the task;
human learning, though slow, may be close to optimally efficient.
I must also enter another caveat because you'll ask me, "What were you
saying in that talk you gave two months ago? Why were you talking about
learning?" The caveat is: Even in a world in which there are lots of computers it
still may be important for us to understand human learning. Artificial intel-
ligence has two goals. First, AI is directed toward getting computers to be smart
and do smart things so that human beings don't have to do them. And second,
AI (sometimes called cognitive simulation, or information processing
psychology) is also directed at using computers to simulate human beings, so
that we can find out how humans work and perhaps can help them to be a little
belter in their work.
28 CHAPTER 2: WHY SHOULD MACHINES LEARN?
None of the doubts I have just raised about computer learning apply to this
second application of AI. Anybody who is interested in machine learning be-
cause he wants to simulate human learning- because he wants to understand
human learning and thinking, and perhaps improve it can pursue his interest in
good conscience. Hut what about those who have other goals?
2.3 WHAT IS LEARNING?
When I had arrived at this point ami surprised myself by writing down
these notes, I asked myself, "What can we talk about legitimately for the next
three days, other than cognitive psychology?" Dut I looked at the names of the
people who were going to be here and at some of the titles of papers in the
program, and I decided that a good deal of what we were talking about wasn't
really learning anyway, so it was all right.
Let me elaborate on that remark. The term "learning", like a lot of other
everyday terms, is used broadly and vaguely in (he English language, and we
carry those broad and vague usages over to technical fields, where they often
cause confusion. I just saw a notice of a proposed special issue of SIGART,
with a list of kinds of learning. It's a long list, and I'd be astonished if all of
the items on it denote the same thing. Maybe it is just a list of the different
species of learning, but I suspect that it also reflects the great ambiguity of the
term "learning".
2.3.1 A Definition of Learning
The only partially satisfactory definition I've been able to find is that learn-
ing is any change in a system thai allows it to perform belter the second lime on
repetition of the same task or on another task drawn from the same population.
The change should IK more or less irreversible -not irreversible in the sense that
you can't unlearn (although that sometimes is hard, especially unlearning bad
habits) but irreversible in that the learning doesn't go away rapidly and
autonomously. Learning denotes changes in the system that are adaptive in ihe
sense thai they enable, the system to do the \arne task or tasks drawn from the
same population more efficiently ami more effectively the next time.
Since we may want the same task clone over and over and over again,
tuning a system so that it runs very fast is a great thing. Human IK ings seem to
have some tuning capabilities, often called automating task performance. Dut
more often, particularly in the university, we're interested in learning, not so that
the same task can be done over and over again, but so that we acquire the ability
to perform a wide range of tasks (for example, solving problems that appear on
examinations, or performing similar tasks that may occur afterwards in real life).
 29
SIMON
: 2.3.2 Learning and Discovery
There are relations between learning and sonic other activities. For one
thing, learning is related to discovery. By discovery I mean finding new Ihings.
Very little human learning is discovery. Most of what we know somebody told
us about or we found in a textbook. At the very best we acquired it by working
out some very carefully selected exercises, which guided us nicely in the right
direction and provided most of the selective heuristics for our search. There can
be all kinds of learning without discovery, and there usually arc. Most of the
things we know were discovered by other people before we knew them, and only
a few were even reinvented by us.
Nevertheless, there is a relation between learning and discovery, because if
you do discover something and it's good, you'd like to retain it. So, if you have
a discovery system, you would like (somehow or other) to associate a learning
system with it, even a simple memorization and indexing scheme. That doesn't
quite gel us off the hook. If you have a computer that discovers the proof for
Goldbuch's Theorem or the Four Color Theorem, you don't have to have a
separate learning program, for you can simply gel the proof out of the computer
and transport it around on paper in the usual way. Out, it would be very con-
venient if the computer could store the proof so (hat it could be used in sub-
sequent work.
One of the first learning programs for computers was the little learning
routine in the Logic Theorist (LT) (Newell & Simon, 1956). When the Logic
Theorist had the good fortune to prove a theorem in Principia Maihcmaiica it
had the good sense to keep the theorem around. On the next problems, it didn't
start from the axioms alone but could use the new theorem along with the
axioms. It wasn't any great feu! to program this learning program. It did what
we teachers call (pejoratively) "rote learning" just memorizing. LT memorized
only the theorem, not the proof; but giving it the latter capability also would
have been a trivial mailer.
In the Artificial Intelligence literature, the distinction I have been main-
taining here between discovery and learning is not usually observed. That is to
say, a great many machine "learning" systems arc also discovery systems; they
discover new knowledge that they subsequently retain. Most of the skeptical ar-
guments I have. raised about machine learning do not apply to the discovery
process. Hence, I think it quite appropriate that a large part of the research ef-
fort in the domain of "machine learning" is really directed at "machine
discovery". As long as we arc not ourselves confused by the terminology, I do
not even see a strong reason to object to this substitution of terms.
2.3.3 Learning and Understanding Natural Language
So, there's a connection here between learning and discovery. There is
also a connection between learning and.understanding. Understanding includes
the whole natural language problem. In human life (and I'll try later to connect
30 CHAPTER 2: WHY SHOULD MACHINES LEARN?
this up with computers) most of what we learn we get from other people, com-
municated to us in natural language. A good many of the tasks that people have
undertaken for machine learning have involved a natural language front end as an
important part of the task. It is also a very annoying part of the task, eating up
nil of your time and energy when you wish you were doing something else.
2.3.4 Learning and Problem-Solving
Additionally, some things we might call "learning" could also be called
"problem-solving". I've heard "automatic programming" called "learning". The
aim of automatic programming is lo be able lo say the same brief vague Ihings to
a computer you'd say lo a good human programmer in defining a task for him
and lo come out with a program on the other end. What the automatic program-
ming program does is not really learning; it is solving the problem of getting
from ihc sloppy ill-slniclured input statement of the programming problem to a
well-structured program in the programming language. This kind of "learning"
could readily come under the usual heading of "problem-solving".
Nevertheless, traditionally at least, (he tasks of discovery, of natural lan-
guage understanding, and of self-programming have often been intermingled
with, or even identified as. learning tasks. If you want to call it learning you
won't get an argument from me. It really isn't learning but ...
2.4 SOME LEARNING PROGRAMS
I'm going to hack off one step further from my unkind words about
machine learning and look at some "classical" examples ("classical" in the field
of computer science is anything twenty years old) of learning programs, to see
whether they really justify my harsh judgment.
2.4.1 Learning to Play Checkers
The first that I ought to mention is surely Arthur Samuel's checker
program |Samuel, 1959]. Here was a program that, in the morning, wasn't very
much of a checker player. Hut after you switched on its learning process and
gave it games to play and other training exercises, by evening it was a State-
champion-levcl checker player. That is a lot better than any of us could do. So
there's a very impressive example of a learning program going back twenty-five
years.
Let me submit that however fine this program was from an Al standpoint,
it only made sense if we really didn't understand checkers. If Samuel had un-
derstood checkers well, he could have put the final evaluation function in right at
the beginning. (You may recall (hat he used two kinds of learning, but the only
one I want lo mention at the moment is tuning the evaluation function for posi-
tions on the basis of outcomes. When gotnl Ihings happened, items that were
.SIMON 31
heavily weighted in the evaluation function got additional weight, and when bad
things happened they lost some of their weight.) If Samuel had known the right
evaluation function at the outset, he would have put it in the program; he would
not have gone through all the learning rigamarole. It cost only one day of com-
puting time, to be sure, but computers were expensive then, even for one day.
It docs make sense to provide for such learning in a task where you don't
know enough to do the fine tuning. We might think of this as an area of
machine learning (or, more accurately, machine discovery) where we can get the
system to behave better than it would behave if we just sat down and
programmed it. Nobody writing chess programs has had this feeling yet. They
all think they know more chess than u computer could acquire just by tuning
itself. As far as I know, none of the successful chess-playing programs have had
any learning ability.
So there arc cases where the computer can learn some things that we didn't
know when we programmed it. But if you survey the fields of Al and
knowledge engineering today, you will find very few cases where people have
had the feeling this could or should be done, or have had any ideas of how to do
it. Nevertheless, this potential application of learning procedures is certainly one
further qualification on my general stricture against such programs.
I've already mentioned learning by the Logic Theorist, but that was just
convenience, unless LT had reached the point where it was discovering
genuinely new things. If Doug Lena! had let AM (Lenat, I977J run for another
two hours as I kept telling him he should and it had discovered something
completely new, then the learning would make sense, for you would want to
save what had been discovered.
2.4.2 Automatic Indexing
There's something to be said (again, largely on convenience grounds) for
systems that arc capable at least of learning discrimination nets El'AM nets, if
you like [Fcigenbaum, 1963). If you're building up a big data base and adding
information to it all the time, you want easy access to that information, and so
you want an index. It's a lot more convenient to have the system index itself as
it goes along, than to index it by hand. 1 Or if you're building a large production
system and don't want to search it linearly, you're going to incorporate an indek
in the production system to select the order in which it performs the tests. There
is no difficulty in automating that; we have known for twenty-five years how to
do it. So why not?
So there's some room for learning there. I don't know whether there's
much room for learning research, since the technology of growing discrimination
'Dy "indexing" I mean building up a tree or nciuoric of tests sn that you can access a Uala stoic in
ways other than by linear search.
32 CHAPTER 2: WHY SHOULD MACHINES LEARN?
nets, alias indexes, is already pretty well developed, but someone may find a
great new way of doing it.
2.4.3 Perceptrons
A final "classical" example (this is a negative example to prove my point)
is the whole line of Perception research and nerve net learning [Roscnblatt,
I958|. A Perception is a system for classifying objects (that is, a discovery and
learning system) that computes features of the stimulus display, then attempts to
discriminate among different classes of displays by computing linear additive
functions of these features. Functions producing correct choices are reinforced
(receive increased weight), those producing incorrect choices have their weights
reduced. I have to conclude (and here I don't think I am in the minority) that
this line of research didn't get anywhere. The discovery task was just so horren-
dous for those systems that they never learned anything that people didn't al-
ready know. So they should again strengthen our skepticism that the problems
of Al are to be solved solely by building learning systems.
2.5 GROWTH OF KNOWLEDGE IN LARGE SYSTEMS
In the remainder of my remarks, I would like to focus attention on large
knowledge-based Al systems, particularly systems that can be expected to con-
tinue to grow and accumulate over a period of years of use. We may find in
such systems some reasons to qualify a general skepticism about the role of
learning in applied Al. Medical diagnosis systems like INTERNIST (Poplc, 1977)
and MYCIN [Shortliffc, I976|, and the venerable Dl-NDRM. program [Fcigenbaum
el ul., 19711 are examples of the kinds of systems I have in mind.
There has been attention (as in TI-IKHSIAS (l)avis, 1981] and other such
efforts) to designing an effective programming interface between these
knowledge-based systems and the humans who arc supposed to improve them,
and you can call that learning (or instruction). Most of the work has been aimed
at making the job of the human easier. (Perhaps that's unfair, for it's a mutual
job for the two of them.) So one might think of the man-machine interface as a
gotxl locus for learning research.
2.5.1 The ISAAC Program
To make my remarks more concrete, I would like to discuss for a bit Gor-
don Novak's well-known ISAAC system, which solves English-language college
physics problems of the sorts found in textbooks (Novak, 1977]. Although
ISAAC is primarily a performance or problem-solving program, one can think of
some interesting ways of complementing it with a learning program.
ISAAC has a data bank containing schcmas that describe various kinds of
simple objects that physicists talk about levers, masses, pivots, surfaces, and
 SIMON 33
;: the like. A schema is just what you'd expect a description list of an object
, with slots that you can fill in with information about its characteristics. In ad-
' diliort, ISAAC has, of course, some productions and a control structure.
> When you give ISAAC a physics problem in natural language out of the
physics textbook, it uses its schemas and productions to produce an internal
representation of the problem. The representation is another node-link structure,
which you can think of as a super-schema made up by assembling and instan-
tiating some of the basic schemas. ISAAC will assemble some levers, and some
masses, and a pivot or two, and a surface in the way the problem tells it, and
make a problem schema out of them. At the outset it parses the sentences slat-
ing the problem, using its schemas to extract structure and meaning from them,
and builds its internal representation, a problem schema.
This internal representation contains so much information about the
problem that ISAAC uses a little subsidiary program to depict the problem scene
on a CRT. Of course the real reason ISAAC wants this internal representation is
not to draw a picture on a scope, but to use it to set up an appropriate set of
equations and solve them.
Notice that ISAAC doesn't try to translate the natural language problem
directly into equations, as Bobrow's STUDENT program did for algebra (Bobrow,
1968]. It first builds up an internal representation what I think a physicist
would call a physical representation (a "mental picture") of the situation. It then
uses that intermediate representation to build the equations, which it ultimately
solves. The internal representation does a lot of work for ISAAC because it iden-
tifies the points where forces have to be equilibrated and therefore identifies
which equations have to be set up.
2.5.2 A Learning Extension of ISAAC
We can enlarge ISAAC by adding to it an UNDERSTAND program (Hayes &
Simon, 1974]. Now you're going to say, "Ah ha! ISAAC already has an under-
standing program, because ISAAC can understand the problems it is given." That
is true. But to do this, ISAAC must already have in memory a rich set of
schemas describing physical devices, and it must already have the set of produc-
tions that allow it to organize these schemas into an internal representation. So
ISAAC already knows all the physics it's going to know. While it understands
problems, how about understanding physics! This would require the ability to
use natural language information to construct new schemas and new productions.
This is what the UNDERSTAND program docs not for physics, but for slightly
simpler domains. UNDERSTAND creates, from the natural language, schemas for
the kinds of objects being talked about and their relations. (In fact, Novak is
presently exploring similar lines of investigation.)
What I want to ask about this whole amalgam of ISAAC and UNDERSTAND
is, what is the place here for learning research in AI? (I know what the place is
here for learning research in psychology. I think this is a very important area.
34 CHAPTER 2: WHY SHOULD MACHINES LEARN?
But let's continue to talk about the AI side of it.) If we understand the domain
1 ourselves, if we understand physics, why don't we just choose an internal
representation and provide the problems lo the system in thul internal represen-
tation? What's all this learning and natural language understanding about? Or,
if we still want lo give the system 11 capability of doing the problems in the back
of the textbook, which are in natural language, then lets build Novak's ISAAC
1 system. Why go through all the rigamarole of an UNDERSTAND program to learn
i the schemas and the productions painstakingly instead of just programming
them? Before you launch into such a project as an AI effort (as distinct from a
psychological research project), you have to answer that question.
2.6 A ROLE FOR LEARNING
Since you have listened very patiently to my skeptical challenge to learning
as the road to the future in AI, I think I should own up to one more important
qualification that needs lo be attached to my thesis- a little fragment of the more
complete antithesis that the other papers of this volume develop.
I began by running down the human species emphasizing how stupid we
all are as revealed by our agoni/ingly slow rates of learning. It is just possible
that the complexity of the learning process is not an accident but is, instead, an
adaptive product of evolution. The human brain is a very large collection of
programs thul cumulates over a lifetime or a large part of a lifetime. Suppose
that we were allowed lo open up the lid and program ourselves directly. In or-
der lo write debugged programs, modifications of our present programs, we
would have to learn a lot about the internal code, the internal representations of
the knowledge and skills we already possess.
Perhaps you know how knowledge is organi/cd in your brain; I don't know
how it is organized in mine. As a consequence, I think it would IK exceedingly
difficult for me to create a new, debugged code that would be compatible with
what is already stored. This is, of course, a problem that we already encounter
with our time-shared computers today. As we add new utility programs, or
modify the monitors or operating systems, we encounter all sorts of interactions
that make these modifications cumulatively harder to effect. At best, we encap-
sulate knowledge in hosts of separate programs that can operate independently of
each other, but by the same token, cannot cooperate and share their knowledge
effectively. Old programs do not learn, they simply lade away. So do human
beings, their umlcbuggable programs replaced by younger, possibly less tangled,
ones in other human he ails. But at least until the stale of uiulebuggability is
reached, human programs arc modified adaptively and repeatedly by learning
processes that don't require a knowledge of the internal representation.
It may be that for this kind of system (a human brain or the memory of a
very large lime-shared computing system) the only way to bring about continual
modification and improvement of the program is by means of learning
SIMON 35
procedures that don'l involve knowing the detail of the internal languages and
programs. Il is a salient characteristic of human learning procedures that neither
teacher nor learner has a detailed knowledge of the internal representation or data
or process. It may turn out that there aren't procedures more efficient than these
very slow ones that human beings use. That's just a speculation, but we ought
to face the grim possibilities as well as the cheery possibilities in the world.
Even if we had to accomplish our complex programming in this indirect
way, through learning, computers still would huve a compensation the costless
copying mechanism that is not shared by human beings. Only one computer
would have to learn; not every one would have to go to school.
2.7 CONCLUDING REMARKS
By now you are aware that my case against AI research in learning is a
very qualified case with several important exceptions exceptions you may be
able to stretch until they become the rule. Let me put the matter in a positive
way, and rephrase these exceptions as priorities for learning research. They are
five in number.
1. I would give a very high priority to research aimed at simulating, and
thereby understanding, human learning. It may IK objected that such
research is not AI but cognitive psychology or cognitive science or some-
thing else. I don't really care what it is called; il is of the greatest impor-
tance that we deepen our understanding of human learning, and the AI
community possesses a large share of the talent that can advance us toward
this goal.
2. I would give a high priority, also, to basic research aimed at understanding
why human learning is so slow and inefficient, and correspondingly, at ex-
amining the possibility mat machine learning schemes can be devised that
will avoid, for machines as well as people, some of the tediousness of
learning.
3. I would give a high priority to research on the natural language interface
between computer systems and human users. Again, it does not matter
whclhcr you call il research on learning or research on understanding. We
do want systems, particularly in the knowledge engineering area, in which
we don't have to know the internal language or representation in order to
interact with them. This is especially true, as I have just argued, if the
systems arc to be cumulative over many years.
4. I think there is an important place for research on programming from in-
complete instructions (automatic programming), which is not unrelated to
the preceding item. Giving instructions to a skilled programmer is dif-
ferent from writing the program yourself else why hire the programmer?
It is a very important research question to ask whether we can get the com-
puter to be the skilled programmer.
36 CHAPTER 2: WHY SHOULD MACHINES LEARN?
5. My final priority is research on discovery programs programs thai dis-
cover new things. We may regard discovery itself as a form of learning,
but in addition we will want to give a discovery system learning
capabilities because we will want il to preserve and to be able to use all
the new things it finds.
So now, I guess. I have come full circle, and have made a strong case for
machine learning. Dili I do not think the effort in addressing my initial skep-
ticism has been wasted. Research done in the right area for the wrong reasons
seldom achieves ils goals. To do good research on machine learning, we must
have clear targets to aim at. In my view, the usual reasons given for AI learning
research are too vague to provide good targets, and do not discriminate with suf-
ficient care the learning requirements for people and computers, respectively.
Perhaps the deepest legitimate reason for doing machine learning research
is that, in the long HIM for big knowledge-based systems, learning will turn out
to be more efficient than programming, however inefficient such learning is.
Gaining a deeper understanding of human learning will continue to provide im-
portant clues about what to imitate and what to avoid in machine learning
programs. II* this is true, (hen it follows that among the most important kinds of
learning research to carry out in AI arc those that are oriented toward under-
standing human learning. Mere as elsewhere. Man seems to be the measure of
all things.
ACKNOWLEDGMENTS
This research was supported by Research Grant MH-07722 from the Na-
tional Institute of Mental Health, anil a grant from the Alfred P. Sloan Foun-
dation.
REFERENCES
Bobrow, D. G., "Natural language input ftw a computer problem-solving system," Semantic Infor-
HMtitnt /Yfwx.vfnx. Minsky, M. (Ed.), MIT I'tcss, Cambridge. MA. 1968.
l)avi\, K., "Applications of mola level knowledge to llte coiiMiuclion and use of large knowledge
bases," KnowletlKt-lttixeil Syxtemx in Artijidnl Intellixrncf, Da vis. R. and t.cnat, D. (lids.),
McGraw-llill Hook Company. New Yoik. NY. 1981.
Fcigcnbaum K. A., "The simulation or verbal learning behavior." Conifntit-r.\ and Thought. Feigcn-
biium. P.. A. ami Feldman. J. (lids.). McGruw Hill Hook Company. New York. NY. I96J.
Fcigcnhaum. E. A., lluchanan. H. G. and Ledcrbcrg, J., "On generality and problem solving: A case
study using llie DP.NDRAI. program," Machine lnirllif(fm-e, Mell/er. B. and Michie.
I), (lids ). Edinburgh University Press. Edinburgh. Scotland. 1971.
llaycs, J. R. and Simon, II. A., "Understanding written problem instructions," Knowledge and
Cognition, Gregg. L. W. (Ed.). Lawrence Eribaum Associates. Potomac. MD, 1974.
SIMON 37

Lenal, D. I)., "Automated theory formation in mathematics." PriH-eedingx of the Fifth International
Joint Coherence on Artificial Intelligence, IJCAI, Cambridge. MA. pp. 833-8-42, August
1977.

Newell, A. and Simon, II. A., "The logic theory machine." IKF. Transactions OH Information
Theory. Vol. IT-2. No. 3. pp. 61-79, September 1956.

Novak. O. S.. "Representations of knowledge in program for solving physics problems," Proceed-
ings of the Fifth International Joint Conference on Artificial Intelligence, IJCAI, Cambridge,
MA. pp. 286-291. August 1977.

Popk, II., "The formation of composite hypotheses in diagnostic problem solving," Proceedings of
the Fifth International Joint Conference on Artificial Intelligence, IJCAI. Cambridge, MA,
pp. 1030-1037. August 1977.

Rosenblalt. F., "The pcrccplron: A probabilistic model for information storage and organization in
the brain." Psychological Review. Vol. 65. pp. 386-407. 1958.

Samuel, A. L., "Some studies in machine karning using the game of checkers," IBM Journal of
Research and Development. No. 3, pp. 210-220, 1959.

Shorlliffe, E.. Computer-Based Medical Consultations: MYCIN, American Flsevier Publishing Com-
pany. New York, NY. 1976.